 AI at the edge provides incredible values for application like predictive maintenance but implementing it usually in practice comes with great challenges. In fact, many customers with worked with have been facing these following issues. They have a fleet of motors, they would like to create a device that is able to monitor their motors and learn their behaviors and be alerted when these motors start showing the first signs of wear before a critical failure actually happens. But they don't know where to start. They think they need to collect massive amounts of data and hire data scientists or machine learning experts. They believe they need to build a machine learning model from scratch and train it on the cloud and that it will take lots of time, lots of effort, energy and money. However, with Nano-Edge AI we show them an alternative way to develop edge AI devices in short times with no prior machine learning expertise and only little qualified data. To showcase Nano-Edge AI algorithms abilities, here's a simple anomaly detection setup. We have a motor which is piloted by a ST motor control board where we can easily create two types of anomalies, shaft misalignment and magnetic friction. Here we will try to detect these anomalies by monitoring the vibration pattern on the motor using the ST eval Proteus 1 evaluation board placed on top of the motor. First for this we'll use the three axis accelerometer of the board. Inside the Proteus board microcontroller we've placed a Nano-Edge AI library that will be able to learn the motors normal state and then detect any drift from this initial state as an anomaly. Note that here no sensor data ever leaves the microcontroller. It is directly consumed by the Nano-Edge AI algorithms which are isolated from the eyesight world and then discarded. This Proteus board besides its small form factor and industrial grade sensors and dedicated machine learning core also comes with Bluetooth low energy capabilities and also a companion mobile app for device settings and easy data view. We'll use this mobile app to display Nano-Edge AI's learning status and inference results. Note that here the Nano-Edge AI machine learning algorithms are not free trained. This is very important. They will dynamically learn the vibration behavior for this specific motor from within the microcontroller and then incrementally build the model's knowledge from scratch and this is game changing. Now let's put that into action. First we place the setup in a normal state and reset the board. So the board will be initializing and the learning will be starting. This board is programmed to learn the normal state first, something like 20 or 30 iterations and then the user switches to detection or inference mode where it will compare what it sees to what has been learned and return a similarity score or a percentage. Now that this learning phase is over we see that the inference result of course is 100% because everything is perfectly normal. It is similar to what has been learned. Let's try to see what happens when we introduce anomalies. For example, we can introduce some shaft misalignment and here we see that this anomalies immediately picked up by the Nano-Edge AI algorithms meaning that the similarity score has dropped significantly. Now if we remove this misalignment naturally we go back to normal. Now if we introduce another type of anomaly which is magnetic friction instantly we see the response from the algorithms and we detect anomalies again and when we remove the friction the anomalies disappear and we're back to normal. What is important to understand here is that because this Nano-Edge AI library has the ability to learn directly from the field from scratch it is incredibly adaptable. It means that the exact same algorithm can be deployed to a fleet of motors and it will train differently taking into account each motors little specificities due to their age or their maintenance status or they changing environmental conditions. Now let's see how easy and fast it is to implement this in practice using Nano-Edge AI Studio with no machine learning expertise. Nano-Edge AI Studio is a tool that runs on your desktop and that takes in as input raw sensor data and uses this input to build the best possible machine learning library which is a combination of signal pre-processing and choice of a type of machine learning model. First we select the target which is here of course the Proteus board and the sensor type. Then we import signal examples representing the normal state of our motor on the setup and this is raw accelerometer data logged with the data logger that we just created seconds ago. Then we do the same with some examples of anomalies that could happen on our system and we use this little little data set that we just created to run a benchmark. During this benchmark millions of potential algorithm combinations are being tested until the best candidate library is found. Once we have our best candidate our best combination of algorithms we can use this emulator to test it against new data without having to compile anything or flash any code. And finally once we're happy with the result we can deploy it on a microcontroller. So here the output is a pre-compiled static library in C language it's a .a format. It's ready to be implemented into any C code and flashed onto your target device. Thank you so much for being here today with us. If you would like to hear more about our AI solutions at ST please go to stm32ai.st.com and here you will find many exciting use cases that you can build using any of our AI software solutions. Thank you.